Add ability to bypass profiling entirely by making ProfilingSettings optional

[elliottslaughter]

• Most places now take std::optional<ProfilingSettings>
• The meaning of std::nullopt in this case is "don't profile, just run the kernel exactly once"
• The reason to bypass profiling is that, for GPU tasks, profiling forces a wait on the GPU stream, which is unacceptable in production code
• Specific cases:
  • local-execution: profiling is optional
  • cost estimator: profiling is mandatory
  • realm-execution: profiling has been removed entirely; we never actually reported the results and Realm has better ways of profiling that don't involve blocking

---

This change is